________________________________
A Hybrid Genetic Algorithm for Multi- Depot and Periodic Vehicle Routing Problems
Thibaut Vidal
Teodor Gabriel Crainic Michel Gendreau Nadia Lahrichi Walter Rei
July 2010
CIRRELT-2010-34
G1V 0A6 Bureaux de Montréal : Bureaux de Québec : Université de Montréal Université Laval
C.P. 6128, succ. Centre-ville 2325, de la Terrasse, bureau 2642 Montréal (Québec) Québec (Québec)
Canada H3C 3J7 Canada G1V 0A6
Téléphone : 514 343-7575 Téléphone : 418 656-2073 Télécopie : 514 343-7121 Télécopie : 418 656-2624
www.cirrelt.ca
Thibaut Vidal1,2, Teodor Gabriel Crainic1,3,*, Michel Gendreau1,4, Nadia Lahrichi1,3, Walter Rei1,3
1 Interuniversity Research Centre on Enterprise Networks, Logistics and Transportation (CIRRELT)
2 Department of Computer Science and Operations Research, Université de Montréal, P.O. Box 6128, Station Centre-ville, Montréal, Canada H3C 3J7
3 Department of Management and Technology, Université du Québec à Montréal, P.O. Box 8888, Station Centre-Ville, Montréal, Canada H3C 3P8
4 Department of Mathematics and Industrial Engineering, École Polytechnique de Montréal, P.O.
Box 6079, Station Centre-ville, Montréal, Canada H3C 3A7
Abstract. We propose an algorithmic framework that successfully addresses three vehicle routing problems: the multi-depot VRP, the periodic VRP, and the multi-depot periodic VRP with heterogeneous capacitated vehicles and constrained route duration.
The meta-heuristic combines the exploration breadth of population-based evolutionary search, the aggressive-improvement capabilities of neighborhood-based meta-heuristics, and advanced population-diversity management schemes. Extensive computational experiments show that, the method performs impressively, in terms of both solution quality and computational efficiency. It particular, it either identifies the best known solutions, including the optimal ones, or identifies new best solutions for all currently available benchmark instances for the three problem classes.
Keywords. Multi-depot multi-period vehicle routing problems, hybrid populations-based meta-heuristics, adaptive population diversity management.
Acknowledgements. Partial funding for this project has been provided by the Natural Sciences and Engineering Research Council of Canada (NSERC), through its Industrial Research Chair and Discovery Grants programs, and by the partners of the Chair, CN, Rona, Alimentation Couche-Tard and the Ministry of Transportation of Québec, and by the Fonds québécois de la recherche sur la nature et les technologies (FQRNT) through its Team Research Project Program.
Results and views expressed in this publication are the sole responsibility of the authors and do not necessarily reflect those of CIRRELT.
Les résultats et opinions contenus dans cette publication ne reflètent pas nécessairement la position du CIRRELT et n'engagent pas sa responsabilité.
_____________________________
* Corresponding author: Teodor.Gabriel@cirrelt.ca
Dépôt légal – Bibliothèque nationale du Québec, Bibliothèque nationale du Canada, 2010
© Copyright Vidal, Crainic, Gendreau, Lahrichi, Rei and CIRRELT, 2010
1 Introduction
Vehicle Routing Problem (VRP) formulations are used to model an extremely broad range of issues in many application fields, transportation, supply chain management, production planning, and telecommunications, to name but a few (Golden et al., 2002;
Crainic, 2008; Hoff et al., 2010). The mathematical structure of most VRP formulations is generally simple but only deceivingly so, most problems of interest being NP-Hard and yielding integer programming (IP) combinatorial formulations presenting significant methodological challenges. Not surprisingly, starting with the initial introduction of the VRP by Dantzig and Ramser (1959), routing problems make up an extensively and continuously studied field, as illustrated by numerous conferences, survey articles (e.g., Christofides et al., 1979; Bodin et al., 1983; Fisher, 1995; Desrosiers et al., 1995; Powell et al., 1995; Gendreau et al., 2002; Laporte and Semet, 2002; Br¨aysy et al., 2004; Br¨aysy and Gendreau, 2005a,b; Cordeau et al., 2005, 2007; Br¨aysy et al., 2008a,b; Francis et al., 2008; Laporte, 2009), and books (Toth and Vigo, 2002; Golden et al., 2008).
Surveying the literature one notices, however, that not all problem classes have re- ceived an equal nor adequate degree of attention. This is the case for the problems with multiple depots and periods, the recent contributions to the periodic variant not altering the statement. A second general observation is that most methodological developments target a particular problem variant, the capacitated VRP or the VRP with time win- dows, for example, very few contributions aiming to address a broader set of problems, the Unified Tabu Search (Cordeau et al., 1997, 2001) being a notable exception. This also applies to the classes of problems targeted in this paper.
Our objective is to contribute toward addressing these two challenges. We propose an algorithmic framework that successfully addresses three VRP variants: the multi-depot VRP,MDVRP, the periodic VRP,PVRP, and the multi-depot periodic VRP,MDPVRP, with heterogeneous capacitated vehicles and constrained route duration. The literature on these problems is relatively scarce despite their relevance to many applications, e.g., raw material supply (Alegre et al., 2007), refuse collection (Beltrami and Bodin, 1974;
Russell and Igo, 1979; Teixeira et al., 2004), food collection or distribution (Carter et al., 1996; Golden and Wasil, 1987; Parthanadee and Logendran, 2006), and maintenance operations (Blakeley et al., 2003; Hadjiconstantinou and Baldacci, 1998).
We propose a meta-heuristic that combines the exploration breadth of population- based evolutionary search, the aggressive-improvement capabilities of neighborhood-based meta-heuristics, and advanced population-diversity management schemes. The method, that we name Hybrid Genetic Search with Adaptive Diversity Control (HGSADC), per- forms impressively, in terms of both solution quality and computational efficiency. It particular, HGSADC either identifies the best known solutions, including the optimal ones, or identifies new best solutions for all currently available benchmark instances for the three problem classes
To sum up, the main contributions of this article are: 1) A new meta-heuristic that is highly effective for three classes of important vehicle routing problem classes, the MD- VRP, the PVRP, and the MDPVRP. It equals or outperforms the current best methods proposed for each particular class and requires a limited computational effort. 2) New population-diversity management mechanisms to allow a broader access to reproduction, while preserving the memory of what characterizes good solutions represented by the elite individuals of the population. In this respect, we revisit the traditional survival- of-the-fittest paradigm to enhance the evaluation of individuals by making it rely on both fitness and diversity (distance-to-the-others) measures. Our empirical studies show this mechanism not only efficiently avoids premature population convergence, but also outperforms traditional diversity management methods relative to the general behavior of the solution method. 3) An efficient offspring education scheme that integrates key features from efficient neighborhood search procedures, e.g., memories and granular tabu search concepts.
The paper is organized as follows. Section 2 states the notation and formal definition of the three classes of VRPs we address, while the relevant literature is surveyed in Section 3. The proposed meta-heuristic is detailed in Section 4, its performances are analyzed in Section 5, and we conclude in Section 6.
2 Problem Statement
We present the problem description for the MDVRP, PVRP, and MDPVRP problem classes, together with the notation we use in this paper, and the problem transformation, MDPVRP to a generalized PVRP, which supports the following developments.
The basic Capacitated Vehicle Routing Problem (CVRP) can be defined as follows.
LetG= (V,A) be a complete graph, vertexv0 ∈ V representing the depot where vehicles and the product to be distributed are kept, each other vertex in V\{v0} standing for one of the n customers to be served from the depot. Each customer i, i= 1, . . . , |V| is characterized by a non-negative demandqi and a service durationdi. Arcsaij ∈ A, i, j ∈ V represent the possibility to travel from vertexi to vertexj at a duration cost of cij. A fleet of m identical vehicles with capacity Q is located at the depot, the total duration of the route performed by each vehicle, computed as the total travel and service time required to serve the customers on the route, being limited at D. The goal is to design a set of m vehicle routes servicing all customers, such that vehicle and route constraints are respected, and the total route duration is minimized.
Several depots, d, are available to service customers in the Multi-Depot VRP, m rep- resenting the number of vehicles available at each depot. A time dimension is introduced in the Periodic VRP as route planning is to be performed over a horizon of t periods.
Each customer i is characterized by a service frequency fi, representing the number of visits to be performed during the t periods, and a list Li of possible visit-period com- binations, called patterns. The PVRP aims to select a pattern for each customer and construct the associated routes to minimize the total cost over all periods. Finally, the Multi-Depot Periodic VRP extends the two previous problem settings, the selection of a depot and visit pattern for each customer being required to construct optimal routes.
It is important to note that a linking constraint is implicitly defined, services in differ- ent periods to the same customer being required to originate at the same depot. The MDPVRP reduces to a PVRP when d= 1 and to a MDVRP when t= 1.
Cordeau et al. (1997) proposed a formulation of a generalized PVRP with vehicle and period-dependent routing costs. The authors showed that this formulation includes the MDVRP as a special case, by associating a different period to each depot, such that each customerihas a frequencyfi = 1 and may be served any period. We follow this lead and show that a MDPVRP with t days and d depots can be transformed into a generalized PVRP, by associating a period in the new problem to each (period, depot) pair in the MDPVRP.
LetP be a MDPVRP instance, with periods{0, . . . , t−1}, depots{0, . . . , d−1}andm vehicles per depot. Letibe a customer with frequencyfi and pattern listLi containingp patternsLi ={{p11, . . . , p1fi}, . . . ,{pp1, . . . , ppfi}}. We now define the problemP0 which has t0 =t×d periods andmvehicles available at each period. For each customer iin the new problem,fi0 =fi and the pattern listL0i containsd×ppatterns defined by Equation 1.
L0i = [
a∈ {0, . . . , d−1}
b ∈ {1, . . . , p}
{pb1+a×t, . . . , pbfi+a×t} (1)
where the firstt periods are associated with depot 0, the next are associated with depot 1, and so on. Travel costs (durations) are adjusted to take into account that vehicles operating in period l in the new general PVRP originate from depot l/d.
We rely on this transformation in the development of the proposed methodology and work on (depot, pattern) couples. We thus transform a problem with several attributes into a problem with less attributes, which thus becomes somewhat easier to address. Of course, the method must be computationally efficient to deal with the increased number of periods and the corresponding increase in problem dimension. As the computational results displayed in Section 5 show, we achieve both these goals.
3 Literature Review
This section provides a brief literature review of contributions for the PVRP, the MDVRP, and the MDPVRP. The purpose of this review is twofold. First, to present the most recently proposed meta-heuristic algorithms, particularly population-based ones, for the considered problems. Second, to distinguish the leading solution approaches for the three problem settings.
Population and neighborhood-based meta-heuristics have been proposed for the PVRP.
With respect to the former, Drummond et al. (2001) proposed an island-based parallel evolutionary method, which evolves individuals representing schedules (patterns), the fitness of each individual being obtained by constructing routes for each period with the Clarke and Wright (1964) savings heuristic. Alegre et al. (2007) proposed a scatter search procedure designed especially for PVRPs with a large numbers of periods. As in Drummond et al. (2001), the core of the method is dedicated to the improvement of visit schedules, while a neighborhood-based improvement procedure is used to design routes for each period. Contrasting with the two previous methods, Matos and Oliveira (2004) proposed an ant colony optimization (ACO) approach that first optimizes routes, then schedules. The PVRP is first transformed into a a large VRP containing each customer as many times as given by its frequency and addressed by an ACO method. The problem of distributiing the resulting routes among periods is then solved as a graph coloring prob- lem, with occasional changes in customer patterns to progress toward a feasible PVRP solution. In a final step, an ACO method is used to optimize the plan for each period separately.
Until recently, however, the most successful contributions to this problem were based on the serial exploration of neighborhoods. The local search approach of Chao et al.
(1995) was the first to use deteriorating moves to escape from poor local optima. It also to temporarily allowed relaxation of vehicle-capacity limits to enhance the exploration of the solution space. The tabu search proposed by Cordeau et al. (1997) introduced an innovative guidance scheme, which collects statistics on customer assignments to periods and vehicle routes in order to penalize recurring assignments within the solutions obtained and, thus, gradually diversify the search. For a long period of time, this method stood as the state of the art solution approach for both the PVRP and the MDVRP, as well as, in its Unified Tabu Search version (Cordeau et al., 2001), for a number of other VRP variants. It has only been outperformed recently by the Variable Neighborhood Search of Hemmelmayr et al. (2009), which is built upon various well-known VRP neighborhoods, e.g., the string relocate, the swap, and the 3-opt. Finally, one should notice the VNS algorithm with multilevel refinement strategy of Pirkwieser and Raidl (2010), particularly tailored for large-size instances.
We are aware of only two evolutionary approaches for the MDVRP, both taking ad- vantage of geometric aspects within the problem. Thus, Thangiah and Salhi (2001)
represented solutions as circles in the 2D space of depot and customers, which is closely related to the idea of customer clustering into routes, whereas Ombuki-Berman and Hanshar (2009) introduced a mutation operator that specifically targeted the depot as- signment to “borderline” customers, i.e., customers that are close to several depots. In this case also, however, neighborhood-based methods, such as the record-to-record local search of Chao et al. (1993), the tabu search algorithms of Cordeau et al. (1997) and Renaud et al. (1996), and the simulated annealing method of Lim and Zhu (2006),proved to be more efficient. To date, the most successful approach for the MDVRP remains the adaptive large neighborhood search (ALNS) method of Pisinger and Ropke (2007), which implements the ruin-and-recreate paradigm with an adaptive selection of its destruction and reparation operators.
In the case of the MDPVRP, most proposed algorithms do not consider all attributes simultaneously, but rather apply a successive-optimization approach. Thus, the method developed by Hadjiconstantinou and Baldacci (1998) starts by first assigning all cus- tomers to a particular depot. Given these a priori assignments, customer visits are then successively inserted among available periods to obtain feasible visit combinations. The depot-period VRP subproblems obtained are then separately solved using a tabu search algorithm. Finally, a phase that attempts to improve the solution by modifying the pe- riod or depot assignments through customer interchanges is applied. The overall solution strategy then repeats this sequence of heuristics for a fixed number of iterations. Other such approaches were proposed by Kang et al. (2005) and Yang and Chu (2000), where schedules for each depot and period are first determined, followed by the design of the corresponding routes.
We are aware of only two methods that aim to address problems similar to the MDPVRP as a whole. Parthanadee and Logendran (2006) implemented a tabu search method for a complex variant of the MDPVRP with backorders. The authors also study the impact of interdependent operations between depots, where the depot assignment of a customer may vary according to the periods considered. Significant gains are reported on small test instances when such operations are applied. Crainic et al. (2009b,a) intro- duced the Integrative Cooperative Search (ICS) framework, which relies on the problem decomposition by attributes, concurrent resolution of subproblems, integration of the elite partial solutions yielded by the subproblems, and adaptive search-guidance mecha- nisms. The authors used the MDPVRP with time windows to illustrate the methodology with very promising results, but no results are available for the problems addressed in this paper. Moreover, ICS targets complex problem settings and we provide a simpler way to treat the MDPVRP.
A number of exact methods were also proposed for one or another of the problems we address. Noteworthy are the recent contributions of Baldacci and Mingozzi (2009) and Baldacci et al. (2010) addressing the MDVRP and the PVRP. Exact methods are limited in the size of instances they may handle, but these particular approaches have
proven quite successful in solving to optimality several instances that are used as a test bed for the algorithm we propose.
This brief review supports the general statement made previously that no satisfac- tory method has yet been proposed for the three problem settings. Furthermore, the contributions to the MDPVRP literature are very scarce, those addressing all the prob- lem characteristics simultaneously being scarcer still. Most solution methods proposed address the periodic and multi-depot VRP settings, with neighborhood-based methods yielding, until now, the best results on standard benchmark instances. However, evolu- tionary methods have proven recently to be efficient on the standard VRP (Prins, 2004;
Mester and Braysy, 2007; Nagata and Br¨aysy, 2009) and on a number of other VRP vari- ants, as underlined in a review on evolutionary methods for the VRPTW (Br¨aysy et al., 2004). Noteworthy is the contribution of Prins (2004) for the VRP, which introduced an important methodological element, namely the solution representation for the VRP as a TSP tour without delimiters along with a polynomial time algorithm to partition the sequence of customers into separate routes. This approach was later applied by (La- comme et al., 2005; Chu et al., 2006) to the periodic capacitated arc routing problem, which shares a number of common characteristics with the PVRP. We adopt this solution representation for the population-based method we propose to efficiently address the pe- riodic and multi-depot problems, as well as the MDPVRP as a whole. This methodology is described in the next section.
4 The Hybrid Genetic Search with Adaptive Diver- sity Control Meta-heuristic
The hybrid meta-heuristic we propose is based on the Genetic Algorithm (GA) paradigm introduced by Holland (1975), but includes a number of advanced features, particularly in terms of generation of new individuals, offspring education, and population management, which contribute to its originality and high performance level.
The method evolves a population of individuals, representing feasible and unfeasible solutions, through successive application of a number of operators to select two parent so- lutions, combine them, yielding a new individual, and enhance this offspring. We identify the latter operator as offspringeducation. It includes a series of solution transformations proved to be efficient in neighborhood-based meta-heuristics for routing problems, as well as a feasibility repair procedure.
Feasible and unfeasible solutions are kept in two separate subpopulations, managed to both evolve toward high-quality solutions and maintain a high level of diversity among individuals. This is performed not only when selecting the surviving individuals for
the next “generation”, which is rather standard GA methodology, but also through an evaluation mechanism of individuals combining traditional fitness and a representation of the particular contribution an individual makes to the diversity of the gene pool. This fitness-diversity evaluation mechanism is shown to play an important role in the overall performance of the proposed methodology.
The general structure of the proposedHybrid Genetic Search with Adaptive Diversity Control (HGSADC) meta-heuristic therefore is
• Initialization (Section 4.6);
• Repeat
– Selection of parent solutions and generation of offspring (Section 4.4);
– Offspring education; If unfeasible, repair with probability Prep (Section 4.5);
– Adjustment of parameters enforcing penalties for violating feasibility condi- tions (if necessary) (Section 4.6);
– Diversification if best solution not improved for Itdiv consecutive iterations (Section 4.6);
– Culling of any sub-population that reaches its maximum number of individuals (Section 4.6);
• Until stopping conditions are satisfied, i.e., eitherItN I iterations without improve- ment of the best feasible solution, or the time limit Tmax is reached.
We initiate the description with the definition of the search space (Section 4.1) and the representation of the individuals (Section 4.2). We then proceed with detailed discus- sions of the individual evaluation procedure (Section 4.3), the selection, crossover, and education operators, and the population management mechanism.
4.1 Search space
It is well-known in meta-heuristic literature that allowing a controlled exploration of unfeasible solutions may enhance the performance of the search, which may more easily transition between structurally different feasible solutions.
A number of constraint relaxations have been proposed in the VRP literature. Thus, for example, Prins (2004) and Lacomme et al. (2005) relax the fleet-size constraints in their GA-based methods, while Gendreau et al. (1994) and Cordeau et al. (1997) relax vehicle capacity and maximum travel time limits in their tabu search algorithms. We
favor the second approach even though our solution representation is closer to that of the first authors. Indeed, a solution with too high a number of vehicles may require sophisticated and computationally costly route-reduction methods. Moreover, it is much easier in the second case to introduce dynamically self-adjusting penalties that gradually guide the search into and out of the unfeasible-solution zone, while still allowing the use of most local search moves used in the VRP literature to modify and enhance routes. We therefore define the search space as the set of feasible and unfeasible solutions, the latter being obtained by relaxing the limits on vehicle capacities and maximum route travel time.
Two penalty parameters are defined, σQ for vehicle load excess and σC for route travel time excess. Given a route r with load q(r) and travel time t(r), Equation 2 defines its penalized cost φ(r). The total cost of a solution is then computed as the sum of the penalized costs of all routes and is used in the computation of the fitness of the individuals relative to the current population.
φ(r) =t(r) +σC×max(0, t(r)−D) +σQ×max(0, q(r)−Q) (2) The penalty parameters σQ and σC are dynamically adjusted during the execution of the algorithm, to favor the generation of “naturally” feasible individuals, that is, individuals that are feasible following education without requiring to call on the repair procedure. LetξREFbe the given desired proportion of naturally feasible individuals, and ξQ andξC the proportion in the last 100 generated individuals of naturally feasible solu- tions with respect to vehicle capacity and route duration, respectively. The adjustment is then performed every 100 iterations (the rules are identical for the duration-constraints parameter ξC):
• if ξQ≤ξREF−0.05, then σQ =σQ∗1.2;
• if ξQ≥ξREF+ 0.05, then σQ=σQ∗0.85.
We do not consider the initial penalty parameter values as particularly critical, be- cause the dynamic adjustment drives them towards suitable values, as long as the initial penalties are of an order of magnitude comparable to the objective function. We therefore set the initial values for these parameters to σC = 1 and σQ = ¯c/q, where ¯¯ c represents the average distance between two customers and ¯q is the average demand.
4.2 Solution representation
An individual I in HGSADC is defined by three “chromosomes” related to its customer schedule, depot allocation, and corresponding routes. Thepattern chromosome,PI, thus registers for each customer i the list of days of visitλI[i], which must correspond to one of its pre-defined patterns. The depot chromosome, DI, contains the depot allocation dI[i] of each customer i. Finally, the route chromosome, RI, contains t×d sequences of customers without trip delimiters, the sequence sI[k, l] corresponding to the customers served from depot k during period l. The route chromosome corresponds to a giant TSP tour representation as introduced by Prins (2004) and illustrated in Figure 1 for an instance with two periods and two depots. The corresponding solution is encoded as four sequences of visits without delimiters, one for each couple (period, depot). TheSplit algorithm is then used on each sequence to find the optimal segmentation into routes and compute the fitness of the individual.
Figure 1: Giant TSP representation of the routes of a MDPVRP solution
TheSplit algorithm described in Prins (2004) reduces the problem of finding the route delimiters to a shortest path problem on a particular graph and operates inO(n2), where n is the number of customers in the sequence. Chu et al. (2006) extended it to account for fleet-size limitations. We further generalize the algorithm to work with penalized unfeasible solutions.
Let Si be the customer in position i of the visiting sequence. Define an auxiliary graph H = (V,A), where V contains n+ 1 nodes indexed from 0 to n. For each pair i < j, an arc (i, j) represents the trip ri+1,j starting from the depot, visiting customers Si+1 to Sj and coming back to the depot. Its associated travel time and load are given by Equations (3) and (4). The arc (i, j) is included in A only if the total load of the corresponding trip does not exceed a value Qmax, which we set at Qmax = 2Q to avoid solutions that are too far from feasibility and reduce the number of arcs. The associated
arc cost is noted φ(ri+1,j).
q(ri+1,j) = P
l=i+1,j
qSl (3)
t(ri+1,j) = c0,Si+1 + P
l=i+1,j−1
(dSl+cSl,Sl+1) +dSj +cSj,0 (4)
An optimal segmentation of the giant tour consists in identifying a minimum-cost path from 0 to n in H containing less than m edges, where m is the number of vehicles available per period. This minimum-cost path can be computed in m iterations of the Bellman-Ford algorithm (see Cormen et al., 2001, for an implementation), each iteration executing in O(n2). When the demand or the distance between customers is “large”, it is possible to impose a bound b on the number of valid trips ending at a given customer i. Thus the complexity of an iteration becomesO(n×b), and the Split algorithm works inO(m×n×b).
Figure 2: Illustration of a Split graph and shortest-path solution
Figure 2 illustrates the Split algorithm on a sequence of 5 customer visits S1 to S5. The first graph presents the problem data: cost of each arc and the customer demands (in bold). In this example, the vehicle capacity is set to Q = 6, thus Qmax = 12, the maximum route duration to D = 150, each customer i has an identical service time di = 10, and the penalty parameters are σQ = 10 and σC = 1. The corresponding graph H displays on each arc the associated route cost including penalties. Thus, for example, the route servicing customers S3,S4, andS5 has a cost of 165 + 20 + 15, the penalties of 20 and 15 corresponding to its load excess (two units) and extra duration, respectively.
The optimal solution of the minimum-cost path problem, with a cost of 260, is made up
of the three following routes: route 1 visits S1, route 2 visits S2, S3, and S4}, and route 3 visits S5. Notice that in the actual implementation of the algorithm, there is no need to explicitly build the graphH (Chu et al., 2006).
4.3 Evaluation of individuals
A major difficulty in population-based algorithms is to avoid premature convergence of the population. The danger of premature convergence is more acute in hybrid methods, where the double impact of education and natural selection heavily favors individuals with good characteristics. Several methods have been proposed to address this issue, e.g., eliminating identical solutions, the so-called clones (Prins, 2004), or imposing inter- individual distance constraints when inserting individuals into the population (S¨orensen and Sevaux, 2006; Lozano et al., 2008).
We include some of these ideas into our population-management mechanism (Section 4.6). We also propose, however, a mechanism that addresses the population-diversity issue continuously, during parent and survivor selection and, thus, right at the level of the admittance to the offspring-creation process. Indeed, our very evaluation function accounts not only for the fitness of an individual, but also for its contribution to the population diversity, aiming to equilibrate the drive for the best individual (elitism), and the possible loss of information usually associated to this drive. It is thus an adaptive mechanism to control the diversity of the population while still aiming for an elitist behavior of the meta-heuristic.
We define the diversity contribution ∆(I) of an individual I as the average distance to its nclose closest neighbors, computed according to Equation (5). Several distances measures were tested in the experiments leading to this final algorithm. A normalized Hamming distanceδH(I1, I2), based on the differences between the service schedules and depot assignments of two customers I1 andI2 appeared the most adequate for the multi- depot, period routing problems we address. This distance is computed according to Equation (6), where 1(cond) is a valuation function that returns 1 if the condition cond is true, 0, otherwise.
∆(I) = 1 nclose
X
J∈Nclose
δH(I, J) (5)
δH(I1, I2) = 1 2n
n
X
i=1
(1(λI1[i]6=λI2[i]) +1(dI1[i]6=dI2[i]) (6)
Individuals are sorted by fitness and diversity contribution. Let f it(I) ∈ [0,1] and
dc(I)∈ [0,1] be the normalized ranks of individual I relative to the current population according to each criterion, respectively, where the best individual has rank 0 and the worst has rank 1. TheBiased Fitness functionBF(I) given by Equation (7) establishes a trade-off between fitness and diversity, and it is used to evaluate the quality of an individual during parent (Section 4.4) and survivor (Section 4.6) selections. It depends upon the actual number of individuals in the populationnbIndiv, and a parameternbElit that corresponds to the desired number of elite solutions for which survival is guaranteed from one generation to the next.
BF(I) = f it(I) +
1− nbElit nbIndiv−1
×dc(I) (7)
4.4 Selection and Crossover
We propose a new periodic crossover with insertions (PIX) for the MDPVRP. This op- erator produces a single child C, which inherits characteristics from its two parents P1 and P2. Each parent is selected by a binary tournament. The binary tournament selec- tion consists in picking from the union of the two subpopulations two individuals with uniform probability, and keeping the one with the best Biased Fitness. Both feasible and unfeasible individuals can be selected to undergo crossover in order to lead the research close to the borders of feasibility, where we expect to find optimal solutions.
The crossover operator was designed to transmit good sequences of visits, while en- abling pattern, depot and route recombinations. Another very important property of the operator is its versatility, as it allows both a wide exploration of the search space, by combining genetic material from the parents in nearly equal proportions, and small refinements of a “good” solution, by copying almost the totality of a parent, along with small parts of the other one. The whole process takes place in four steps.
Step 1. Determine period and depot inheritance. The first task of the crossover is to determine for each couple (depot, period) whether the genetic material ofP1, P2 or both is transmitted. Choosing randomly for each depot and period among the three previous cases would not be suitable for the last property stated previously, because when the number of couples becomes large, the amount of information taken from each parent may tend to be balanced because of the law of large numbers. To keep the possibility to focus on one of the parents, we proceed as follows. Two random numbers are first picked between 0 andtd according to a uniform distribution; letn1 and n2 be the smallest and the largest of these numbers, respectively. n1 couples (depot, period) are then selected at random to form the set ΛP1 and n2−n1 of the remaining couples are again randomly picked to form the set ΛP2; the remaining td−n2 couples make up the set Λmix.
Step 2. Copy data from the first parent. For each depotk and period l, if (k, l)∈ΛP1 all visits are copied fromsP1[k, l] intosC[k, l]. If (k, l)∈Λmix, two cutting points αkl and βkl are generated in sP1[k, l], and the corresponding substring is copied intosC[k, l].
Step 3. Copy data from the second parent. For each depotk and periodl selectedin random order, if (k, l) ∈ ΛP1, no visit is copied from sP2[k, l]. If (k, l) ∈ ΛP2, all the visits of sP2[k, l] are successively considered, starting from the beginning. A customer visit i ∈ sP2[k, l] is copied at the end of sC[k, l] if the depot choice dC[i] is undefined or equal to k, and if there exists at least one visit pattern of customer i containing the sub-list λC[i]∪l. Finally if (k, l) ∈ Λmix, we do the same as when (k, l) ∈ ΛP2, but all the visits of sP2[k, l] are considered circularly from βkl.
Step 4. Fill the remaining services. For each customer whose frequency is not satisfied, additional visits are placed in a random fashion as many times as needed inC via a least cost insertion in the sequence, such that the resulting visit list corresponds to an existing pattern, and the same depot is used for all the visits to a given customer.
Figure 3: The PIX crossover after Step 3
An illustration of the crossover operator PIX is given in Figure 3. In this exemple ΛP1 = {(1,1)}, hence (depot 1, period 1) of P1 is copied totally. ΛP2 = {(1,2)}, and for other combinations of periods and depots, C inherits a partial sequence of P1, with additional visits (Λmix ={(2,1); (2,2)}).
4.5 Education
The offspring resulting from the crossover operation undergoes the Split procedure to extract its routes. AnEducation operator is then applied with probabilityPm to improve the quality of the new individual. Education goes beyond the classical GA concepts of random mutation and enhancement through hill-climbing techniques, as it includes
several local-search procedures based on well-known neighborhoods for the VRP. ARepair phase completes the Education operator when the educated offspring is unfeasible.
Two sets of local-search procedures are defined. The nine route improvement (RI) procedures are dedicated to optimize each VRP subproblem separately, whereas the pattern improvement (PI) procedure relies on a quick and simple move to improve the period assignment of the customers by changing their patterns and depots. These local searches are called in the RI, PI, RI sequence.
Route Improvement. Define the neighborhood of a vertex, customer or depot, u as the h×n closest vertexes to u, where h∈[0,1] is a granularity threshold restricting the search to nearby vertexes (Toth and Vigo, 2003). Let v be a neighbor of u, x and y the successors ofuandv in their respective routesT(u) andT(v), respectively. The notation (u1, u2) identifies the route fromu1 tou2. The Route Improvement phase is performed by browsing in a random order each vertex u and each of its neighbors v. For each couple, we evaluate the following moves:
• (M1) If u is a customer visit, remove u and place it after v.
• (M2) If u and x are customer visits, remove them, then placeu and x afterv.
• (M3) If u and x are customer visits, remove them, then placex and u afterv.
• (M4) If u and v are customer visits, swapu and v.
• (M5) If u, x, andv are customer visits, swap u and x with v.
• (M6) If u, x,v, and y are customer visits, swap u and x with v and y.
• (M7) If T(u) = T(v), replace (u, x) and (v, y) by (u, v) and (x, y).
• (M8) If T(u)6=T(v), replace (u, x) and (v, y) by (u, v) and (x, y).
• (M9) If T(u)6=T(v), replace (u, x) and (v, y) by (u, y) and (x, v).
Moves M1 to M6 are applied indifferently on the same or on different routes. M7 is an intra-route move, while the last two are inter-route swaps. The first three moves correspond to insertions, the moves M4 to M6 are swaps, while moves M7 to M9 are known under the names of 2-opt and 2-opt*. It is important to browse the nodes in random order as always visiting a given node before another could lead to bias. As soon as an improvement is found the move is performed. The Route-Improvement phase stops when all possible moves have been successively tried without success.
Pattern Improvement. Letψ(i, k, l) be the minimum insertion cost of customeriin a route from depot k in period l. Let k∗ and p∗ be the depot and pattern, respectively, of
customer iin the current solution. The Pattern-Improvement procedure iterates on each customer i in a random order and computes Ψ(i, k, p) = P
l∈pψ(i, k, l) for each depot k and pattern p∈Li. If Ψ(i, k, p)<Ψ(i, k∗, p∗), then all visits to customer i are removed, and a new visit is inserted in the best location in each sequence corresponding to depot k and period l ∈ p. The procedure stops when all customers have been successively considered without success.
The Pattern-Improvement procedure is significantly faster when the optimal position and insertion cost of each customer is stored for each route. The proportion of routes changed by a move, and the resulting loss of information, is generally small compared to the total number of routes for all depots and periods of the problem. It is also worth noting that, sometimes, the current pattern and depot choices are kept, but a better insertion of the customers is found. The resulting move is then, in fact, a combination of intra-day M1 insertions. This may prove particularly interesting for the exceptional case when the move was not attempted in RI because of proximity conditions. The Pattern- Improvement phase thus fulfills the double role of changing the patterns, and attempting moves between distant vertexes.
A feasible individual yielded by the RI, PI, RI sequence is, of course, inserted into the feasible subpopulation. An unfeasible one is either subjected to the Repair procedure with probability Prep, or inserted into the unfeasible subpopulation. If Repair is successful, the resulting individual is added to the feasible subpopulation.
Repair consists in temporarily multiplying the penalty parameters by 10 and re- starting the RI, PI, RI sequence. When the resulting individual is still unfeasible, penalty parameters are multiplied by 100 and the sequence is started again. This significant increase in the value of the penalties aims to redirect the search toward feasible solutions.
The individual is discarded in case of failure. Computational experiments showed that Repair is crucial to attain good solutions on many tightly constrained instances, e.g., when total vehicle capacity is close to the total demand or the route-duration limit is too tight for the number of available vehicles.
4.6 Population management
The population management mechanism complements the selection-crossover-education operators in identifying the characteristics of good solutions and providing the means for a thorough and efficient search. We propose thus a mechanism with the dual purpose of preserving the accumulated memory of successive selections for the most promising solution characteristics and the diversity of the individuals in the population.
The mechanism manages both the feasible and the unfeasible subpopulations and is made up of three components: the initialization of the populations, a diversification
scheme applied when the search does not seem to advance at a purposeful pace, and the procedure to select the individuals that will survive to the next “generation” when a particular subpopulation reaches its maximum size.
Each subpopulation contains betweenµand µ+λindividuals. To initialize them, 4µ individuals are created by randomly choosing a pattern for each customer and producing for each period the associated service sequence in random order. Each one of these initial individuals undergoes education, and half of those are repaired (if needed). The resulting solutions are stocked in the appropriate subpopulation. If either one of the two subpopulations reaches the maximum sizeµ+λ,λ individuals are discarded as explained later, in the survivor-selection part. At the end of the Initialization phase, one of the two subpopulations can be incomplete, having less thanµ individuals.
During the search, diversification is called whenever Itdiv iterations are performed without improving the best solution. Diversification is performed by retaining only the best µ/3 individuals of each subpopulation, and completing with 4µ new individuals as in the Initialization phase. This process introduces a significant amount of new genetic material, which helps to pursue the search further, even when the population has lost most of its diversity.
The proposed meta-heuristic proceeds by keeping the parents following crossover and having the offspring start to reproduce immediately after education. The populations therefore grow continuously. As the maximum size of µ+λ individuals in one subpop- ulation is reached, it is reduced by discarding λ individuals. The remainingµ survivors provide the next generation. The selection of the survivors aims to preserve the popu- lation diversity in terms of visit patterns, profit from new individuals, and protect an elite. We thus aim to eliminate “clones” and solutions that are bad both in fitness and contribution to diversity, as measured by their Biased Fitness function value.
Let a clone be an individualI2 with the same attributes as another individualI1, i.e., δH(I1, I2) = 0, or the same fitness. The selection of the survivors is then performed as the successive elimination of, first, the clones, and, then, bad solutions, until λ individuals are discarded:
• LetX be the set of all individuals, different from the current best solution, having a “clone”;
• IfX 6=∅, remove I ∈X with maximum Biased Fitness;
• Otherwise, remove I in the subpopulation with maximum Biased Fitness;
• Update the distance measures andX, and repeat.
We can now state an important elitism property. proposition Using the Biased
Fitness function, if an individual I 6∈ X is part of the nbElit best individuals of the subpopulation in terms of fitness, then I will not be removed from the population by the previous survivor-selection procedure.
Proof. Let I be the individual with the worst fitness, thus f it(I) = 1 and BF(I) = f it(I)+(1−nbIndiv−1nbElit )×dc(I)≥1. LetJ be an individual among the best nbElitsolutions in terms of fitness, thus BF(J)≤ nbIndiv−1nbElit−1 + 1− nbIndiv−1nbElit <1. Individual J will not be removed as I has a worst biased fitness.
Figure 4: Illustration of the survivor selection property
Figure 4 represents the fitness and diversity measures of a subpopulation taken from our experimentations after the survivor-selection procedure was run. As stated by the previous proposition, this removal policy calls for nbElitindividuals to be considered an elite. The figure also displays theremoval zone, where individuals with withBF ≥1 can be eliminated according to the Biased Fitness criteria. The procedure favors individuals with excellent fitness and bad diversity over solutions with bad fitness and excellent diversity, as the former are difficult to find. This proved to be a success factor in the experiments leading to the final algorithm.
5 Computational Experiments
Previous work on the PVRP and the MDVRP has led to the development of a large collection of test problems that have been used extensively to provide comparisons be- tween algorithms. Cordeau et al. (1997) describes two sets of instances. The first gathers several problems proposed in previous article, while the second was generated by the authors. The size of these instances range from 50 to 417 customers. Some instances
in the PVRP set are particularly large, involving up to 900 visits to customers over the different periods. Detailed information on these instances is provided in Subsection 5.2.
No benchmark was available for the MDPVRP. We therefore built a set of 10 MD- PVRP instances by merging the PVRP and the MDVRP instances of the second set provided by Cordeau et al. (1997). Each of the 10 MDVRP instances was combined with the PVRP instance with the same number of customers. The number of periods and the patterns were taken from the PVRP instance, the depots from the MDVRP one, and the number of vehicles was fixed to the smallest number such that a feasible solution may be found by HGSADC (and probably by most any other algorithm). The full data sets can be obtained from the authors.
The proposed meta-heuristic was coded in C++, and run on a 2.4 Ghz AMD Opteron 250 CPU workstation. To allow for easier comparisons with previous work, all CPU times reported in this section were converted into their equivalent Pentium IV 3.0 Ghz run times using Dongarra (2009) factors. Before providing detailed computational results, we first explain how we calibrated the algorithm.
5.1 Calibration of the HGSADC algorithm
Similarly to most meta-heuristics, evolutionary ones in particular, HGSADC relies on a set of correlated parameters and configuration choices for its key operators. able 1 provides a summary of the HGSADC parameters, together with the range of values we estimate as appropriate due to either the parameter definition (e.g., probabilities and proportions), conceptual requirements (a local distance measure is assumed to implicate not more than 25% of the population), or values found in the literature (e.g., subpop- ulations sizes). Calibration aims to identify values for these parameters to ensure good performance of the algorithm over a varied set of instances. It is thus often performed by running the algorithm with a number of parameter-value combinations on a limited, hopefully representative, set of instances.
A different approach is offered by the principle of meta-calibration, which aims to optimize the set of parameters of a given algorithm by means of meta-heuristics. Early studies on meta-calibration can be found in Mercer and Sampson (1978). More recently, Smit and Eiben (2009) provide a comparison of parameter-tuning methods, and em- phasize the efficiency of calibrating genetic algorithms by means of a meta-evolutionary algorithm.
We adopted this approach and performed the calibration of the HGSADC algorithm with a meta-evolutionary method, namely, the Evolutionary Strategy with Covariance Matrix Adaptation (CMA-ES) of Hansen and Ostermeier (2001). The source code is available at (http://www.lri.fr/˜hansen/cmaes inmatlab.html). Retrieving the fitness of
a given set of parameters implicates to launch automatically HGSADC on a restricted set of training instances and measure its efficiency. The calibration was thus performed independently for each problem class, with the dual objective of measuring the depen- dency of the best parameter set upon the problem class, and identifyuing an eventual set of parameters suitable for all problem classes considered. The calibration results for each class, along with the final choice of parameter values for HGSADC, are presented in Table 2.
Except for the generation size λ, the optimum set of parameters appears independent of the problem type. We therefore averaged these results to get the final parameter valuess of Table 2, with the exception of the probability to educate a new individual (the education rate Pm). Calibrated education rates are generally very high, with an average value of 0.8. Additional tests indicated similarly good performance as long asPm ≥0.7.
Hence we selected the value Pm = 1, which corresponds to a systematic education of all individuals, and reduces the number of parameters in use. The only parameter that is problem dependent is λ, which is set to 40 for the PVRP, 70 for MDVRP, and 100 for the MDPVRP.
5.2 Results on PVRP and MDVRP instances
For comparison with other algorithms on Cordeau et al. (1997) instances, we report the results of HGSADC with ItN I = 10000, Itdiv = 4000, and Tmax = 10min. The benchmark contains some very large problems with more than 450 visits to customers, for which the population size was reduced by two, and the computation time limit was increased.
Results on PVRP benchmarks are presented in Table 3. The first four columns display the instance identifier and the numbers of customers, vehicles, and periods, respectively.
The results of of HGSADC are shown in Columns 8 and 9 as the average results and CPU times on 10 runs. We compare the performance of HGSADC to that of state-of-the-art methods for the PVRP: the tabu search of Cordeau et al. (1997) (CGL in Column 5), the scatter search of Alegre et al. (2007) (ALP, Column 6), and the average results of 19 runs of the variable neighborhood search of Hemmelmayr et al. (2009) (HDH, Column 7). We indicate in boldface the best average result among the four algorithms for each instance, as well as, in the last two columns, the previous best-known solution (BKS), and the best solution obtained by HGSADC during all our experiments. Optimality has been proved for 14 instances by Baldacci et al. (2010). (Due to numerical approximations, optimality was proved within a 0.02% precision and, hence, it is sometimes possible to find a slightly better solution.) These solutions are marked with *. Several upper bounds have been improved, and the corresponding new state-of-the-art solutions are underlined.
The last two lines provide average measures over the 42 instances: the average per-
Table 1: Parameters of HGSADC
Population and distance measures parameters
µ Population size [5,200]
λ Number of offspring in a generation [1,200]
el Proportion of elite individuals, such that nbElit=el×µ [0,1]
nc Proportion of close individuals considered for distance evaluation, [0,0.25]
such thatnclose =nc×µ Genetic operators parameters
Pm Education rate [0,1]
Prep Repair rate [0,1]
h Granularity threshold in RI [0,1]
Adaptive penalty parameters
ξREF Reference proportion of feasible individuals [0,1]
Table 2: Calibration Results
Parameter PVRP Set MDVRP Set MDPVRP Set Final parameter values
µ 18 24 30 25
λ 33 87 146 40 / 70 / 100
el 0.38 0.45 0.36 0.4
nc 0.24 0.18 0.15 0.2
Pm 0.86 0.86 0.70 1.0
Prep 0.57 0.61 0.33 0.5
h 0.53 0.36 0.35 0.4
ξREF 0.10 0.30 0.20 0.2
centage of error relative to the previous BKS, and the average computation time of each method. Because CGL was tested on a rather old computer, it was hard to make the conversion of the run time using Dongarra (2009) factors. However, Hemmelmayr et al.
(2009) ran the CGL code on a Pentium IV 3.2 Ghz and provided updated computation times.
HGSADC produces high quality solutions, with an average overall deviation of +0.20%
relative to the previous BKS compared to more than +1.40% for the other approaches.
The method we propose yields the best average results on every instance, but one (p06).
It should be noted that, in the case of problem p06, every customer has a frequency of one and can be served on any period. The problem may therefore be addressed as a CVRP. This induces symmetry, which results in many different representations in terms of periods for the same solution. These may then proliferate in the population and can thus explain why the algorithm is less efficient for this particular instance. The average computation time is reasonably short (5.56 min), and is suitable for many operational decisions. Although higher than for the other methods, the run times remain comparable (i.e., in a similar range of values).
It is noteworthy that the average standard deviation per instance obtained by HGSADC is 0.15%, meaning that the algorithm is very reliable. It also performs at a high level in terms of solution quality. All the previous best-known solutions were found, and new reference results were obtained for 20 instances out of 42. The proposed algorithm almost always retrieves the optimal values for the 14 solutions for which the optimal values are known. An average deviation of only 0.09% to the optimum may be observed for these instances.
To observe the behavior of HGSADC when the number of iterations increases, we provide additional results with three termination criteria allowing more computation time: (ItN I, Tmax) = (2.104,30min), and (ItN I, Tmax) = (5.104,1h). Table 4 compares the average results of HGSADC to those obtained by Hemmelmayr et al. (2009) using various numbers of iterations. No running times were reported by Hemmelmayr et al.
(2009) for these experiments, but it seems reasonable to expect them to grow linearly with the number of iterations.
The figures of Table 4 indicate that the proposed algorithm scales well when the available computation time is increased. It should also be pointed out that even long runs of 109 iterations of the HDH algorithm (100 times the number of iterations in the standard HDH method and probably around 300 minutes of computing time) produce inferior average results compared to those obtained by HGSADC in a few minutes.
Results on MDVRP benchmarks of both sets of Cordeau et al. (1997) are displayed in Table 5, where Columns 2 to 4 indicate the numbers of customers, vehicles, and depots, respectively. The results of HGSADC are compared to those of the state-of-the-
Table 3: Results on Cordeau et al. (1997) PVRP instances
Average BKS
Inst n m p CGL ALP HDH HGSADC T(min) prev BKS HGSADC
(1 run) — (10 runs) (10 runs) — (all exp.)
p01 50 3 2 524.61 531.02 524.61 524.61 0.22 524.61* 524.61*
p02 50 3 5 1330.09 1324.74 1332.01 1322.87 0.44 1322.87 1322.87 p03 50 1 5 524.61 537.37 528.97 524.61 0.18 524.61* 524.61*
p04 75 6 5 837.94 845.97 847.48 836.59 1.05 835.26* 835.26*
p05 75 1 10 2061.36 2043.74 2059.74 2033.72 2.27 2027.99 2024.96
p06 75 1 10 840.30 840.10 884.69 842.48 0.89 835.26* 835.26*
p07 100 4 2 829.37 829.65 829.92 827.02 0.88 826.14 826.14
p08 100 5 5 2054.90 2052.51 2058.36 2022.85 2.54 2034.15 2022.47
p09 100 1 8 829.45 829.65 834.92 826.94 1.01 826.14 826.14
p10 100 4 5 1629.96 1621.21 1629.76 1605.22 1.80 1593.45 1593.43
p11 126 4 5 817.56 782.17 791.18 775.84 4.60 779.06 770.89
p12 163 3 5 1239.58 1230.95 1258.46 1195.29 5.34 1195.88 1186.47
p13 417 9 7 3602.76 — 3835.90 3599.86 40.00 3511.62 3492.89
p14 20 2 4 954.81 954.81 954.81 954.81 0.08 954.81* 954.81*
p15 38 2 4 1862.63 1862.63 1862.63 1862.63 0.17 1862.63* 1862.63*
p16 56 2 4 2875.24 2875.24 2875.24 2875.24 0.32 2875.24* 2875.24*
p17 40 4 4 1597.75 1597.75 1601.75 1597.75 0.27 1597.75* 1597.75*
p18 76 4 4 3159.22 3157.00 3147.91 3131.09 0.89 3136.69 3131.09 p19 112 4 4 4902.64 4846.49 4851.41 4834.50 2.26 4834.34 4834.34 p20 184 4 4 8367.40 8412.02 8367.40 8367.40 4.01 8367.40 8367.40 p21 60 6 4 2184.04 2173.58 2180.33 2170.61 0.90 2170.61* 2170.61*
p22 114 6 4 4307.19 4330.59 4218.46 4194.23 4.27 4193.95 4193.95 p23 168 6 4 6620.50 6813.45 6644.93 6434.10 4.29 6420.71* 6420.71*
p24 51 3 6 3704.11 3702.02 3704.60 3687.46 0.32 3687.46* 3687.46*
p25 51 3 6 3781.38 3781.38 3781.38 3777.15 0.59 3777.15* 3777.15*
p26 51 3 6 3795.32 3795.33 3795.32 3795.32 0.33 3795.32* 3795.32*
p27 102 6 6 23017.45 22561.33 22153.31 21885.70 3.52 21912.85 21833.87 p28 102 6 6 22569.40 22562.44 22418.52 22272.60 4.67 22246.69* 22242.51 p29 102 6 6 24012.92 23752.15 22864.23 22564.05 3.86 22543.75* 22543.75*
p30 153 9 6 77179.33 76793.99 75579.23 74534.38 9.99 74464.26 73875.19 p31 153 9 6 79382.35 77944.79 77459.14 76686.65 10.00 76322.04 76001.57 p32 153 9 6 80908.95 81055.52 79487.97 78168.82 10.00 78072.88 77598.00
pr01 48 2 4 2234.23 — 2209.11 2209.02 0.29 2209.02 2209.02
pr02 96 4 4 3836.49 — 3787.51 3768.86 2.49 3774.09 3767.50
pr03 144 6 4 5277.62 — 5243.09 5174.80 7.32 5175.15 5153.54
pr04 192 8 4 6072.67 — 6011.39 5936.16 10.00 5914.93 5877.37
pr05 240 10 4 6769.80 — 6778.00 6651.76 20.00 6618.95 6581.86
pr06 288 12 4 8462.37 — 8461.45 8284.94 20.00 8258.08 8207.21
pr07 72 3 6 5000.90 — 5007.01 4996.14 1.49 4996.14 4996.14
pr08 144 6 6 7183.39 — 7119.61 7035.52 10.00 6989.81 6970.68
pr09 216 9 6 10507.34 — 10259.09 10162.22 20.00 10075.40 10038.43 pr10 288 12 6 13629.25 — 13342.41 13091.00 20.00 12924.66 12897.01
Avg Gap to BKS +1.82% +1.40% +1.45% +0.20%
Avg Time 4.28 min 3.64 min 3.34 min 5.56 min
